SKILL-DISCO Distills Agent Traces into Reusable Procedural Skills.

Autonomous agents often waste computational resources and time by repeatedly solving similar task instances from scratch, leading to redundant reasoning and lengthy execution traces. While previous work has explored workflow reuse and executable skill induction, it has been unclear how to identify scenarios suitable for procedural skills and how to represent shared procedural structures across successful task completions. SKILL-DISCO addresses this by focusing on FSM-defined scenarios, where successful traces can be viewed as paths within an unknown transition graph. It formulates procedural skills as reusable, parameterized control-flow subgraphs. The framework operates through a distillation-and-compilation process: it distills these reusable PFSM subgraphs from successful agent traces and then compiles them into callable, executable, and verifiable procedural skills. Experiments conducted on benchmarks such as ALFWorld and WebArena demonstrate the effectiveness of SKILL-DISCO. The framework consistently improves success rates and reduces the number of agent turns across various benchmarks and model scales. This highlights the significant benefits of representing shared agent experience as structured, reusable execution components.